STRUCTURE is more robust than other clustering methods in simulated mixed-ploidy populations

Analysis of population genetic structure has become a standard approach in population genetics. In polyploid complexes, clustering analyses can elucidate the origin of polyploid populations and patterns of admixture between different cytotypes. However, combining diploid and polyploid data can theoretically lead to biased inference with (artefactual) clustering by ploidy. We used simulated mixed-ploidy (diploid-autotetraploid) data to systematically compare the performance of k-means clustering and the model-based clustering methods implemented in STRUCTURE, ADMIXTURE, FASTSTRUCTURE and INSTRUCT under different scenarios of differentiation and with different marker types. Under scenarios of strong population differentiation, the tested applications performed equally well. However, when population differentiation was weak, STRUCTURE was the only method that allowed unbiased inference with markers with limited genotypic information (co-dominant markers with unknown dosage or dominant markers). Still, since STRUCTURE was comparatively slow, the much faster but less powerful FASTSTRUCTURE provides a reasonable alternative for large datasets. Finally, although bias makes k-means clustering unsuitable for markers with incomplete genotype information, for large numbers of loci (>1000) with known dosage k-means clustering was superior to FASTSTRUCTURE in terms of power and speed. We conclude that STRUCTURE is the most robust method for the analysis of genetic structure in mixed-ploidy populations, although alternative methods should be considered under some specific conditions.